vikalp-090102/NER_Swahili
Domain:
natural language processingRecord type:
model
Creator:
vik
Host:
Named Entity Recognition for Swahili using Inkuba LM Model
# NLP Project: Swahili Named Entity Recognition using InkubaLM-0.4B
## Project Overview
This project fine-tunes the **Lelapa/InkubaLM-0.4B** language model for Named Entity Recognition (NER) on the **Swahili News dataset**. The goal is to improve entity recognition capabilities in Swahili by leveraging **transformers** and **Hugging Face's datasets**.
## Prerequisites
Before running the project, ensure you have **Python 3.13.0** installed on your system.
### Required Libraries
Install the necessary dependencies by running:
```bash
pip install transformers datasets seqeval torch
```
Additionally, you need to authenticate with **Hugging Face**:
```bash
huggingface-cli login
```
## Dataset
We use the **MasakhaNER** dataset for Swahili NER, which is loaded via Hugging Face's `datasets` library:
```python
from datasets import load_dataset
dataset = load_dataset("masakhaner", "swa")
print(dataset)
```
## Model and Tokenization
We fine-tune **Lelapa/InkubaLM-0.4B** using the `AutoTokenizer` and `AutoModelForTokenClassification`:
```python
from transformers import AutoTokenizer
# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained("lelapa/InkubaLM-0.4B", trust_remote_code=True)
tokenizer.pad_token = tokenizer.eos_token
```
## Training
We use Hugging Face's `Trainer` for fine-tuning. The `DataCollatorForTokenClassification` ensures proper token alignment.
```python
from transformers import AutoModelForTokenClassification, Trainer, DataCollatorForTokenClassification
from datasets import Dataset
model = AutoModelForTokenClassification.from_pretrained("lelapa/InkubaLM-0.4B", num_labels=NUM_LABELS)
data_collator = DataCollatorForTokenClassification(tokenizer)
trainer = Trainer(
model=model,
args=training_args,
train_dataset=dataset["train"],
eval_dataset=dataset["validation"],
data_collator=data_collator,
)
trainer.train()
```
## Evaluation
We evaluate the model using the **seqeval** metric:
```python
from seqeval.metrics import classification_report
predictions = …